[zeniheisser]

Original implementation was written as part of the MadtRex implementation, but I had Claude port it to a separate branch to avoid a huge PR that both introduces MadtRex and modifies existing structures.

Changes: explicit inclusion of physics model and FPTYPE in the names of compiled libraries; necessary in MadtRex to differentiate between different models running on the same backend.
UMAMI parameter setters/getters: Implements the functions in the UMAMI interface to set parameters in a loaded library, rather than having to reparse the entire param_card.dat for every library launch. Probably not useful, but it could also be used to run event generation with varying physics parameters (e.g. a scan).

Claude's comments:

• umami_set_parameter / umami_get_parameter: runtime access to independent model parameters (by name or SLHA location) and to parton masses, for the MadMatrix backends (cpu/simd/gpu umami.cc) and the mg7 C++ api; the exporters register every scalar parameter/coupling in a name -> pointer table. Couplings and parameters are registered into separate maps so that FPTYPE=f builds compile (couplings are cxtype, parameters stay double).
• Process and common libraries now carry the model name and FPTYPE (libmadmatrix__.so), so builds for different models/precisions out of the same directory never collide; me_path in subprocesses.json gains an {fptype} placeholder, resolved by mg7 launch.py and gridpack.py from [run] precision.
• madmatrix.mk/madmatrix_src.mk pass FPTYPE through to the common library build and clean targets.